reya results, conclusions, and next steps

agents evolved in reya demonstrated capabilities of exploring, charging towards targets, and shooting at them, when the evolution occours on the same grid. if each genome gets a unique grid, however, their lack of generalizablity almost halts improvement. with a population of 1000 agents, and after 2000 iterations, the best genome only looked around and charged at a target without shooting it.

i would expect that with a large enough population and enough iterations they would learn to shoot, but overall their intelligence seems to not be generalizable. and while this may work if i just wanted to have a result for the sake of having a result, my goal with this project is to attempt to demonstrate an artificial intelligence that is not explicitly "trained" in the same way traditional AIs are and genrealizes quite well, which i failed to reach.

while i had concluded recently that this sort of evolution where the direct output matrix weights are evolved has too large of a search space to evolve efficiently, and have been coming up with genome models that generates a neural network with a high "compression" rate, it was way out of scope with this project, and i was afraid that the models i had thought of were not at all compatible with the output matrices of reservoir computers.

overall i am slightly disappointed with the results, but the conclusions i have drawn from them have pointed me in the right direction.

the project i want to pick up next would use the aforementioned genome models or something similar to attempt to pass as many test from the artifical general intelligence test bed as possible. at least the ones that pretrained models fail at.